Multi-agent space collaborative governance method and system

Abstract

The application relates to the technical field of space governance, and discloses a multi-agent space collaborative governance method, which comprises the following steps: processing multi-source heterogeneous data such as satellite images and sensor readings, establishing cross-type semantic association through geographic space data embedding technology, generating unified structured text after optimizing the satellite image through a vLLM, and synchronizing to a central database; an agent role image is dynamically generated by a central server large language model based on a preset Prompt template, and the generation is independent of fixed rules; then, a target node is selected in an edge-center architecture, a total task is decomposed into subtasks, semantic mapping is established and a matching probability is calculated, and distribution is optimized by means of a reward function; a governance scheme is generated in a perception layer-decision layer-execution layer architecture, and an encoded operation instruction is output; a strategy is updated based on execution feedback, a multistage mechanism is established, roles are automatically redistributed, and governance continuity is ensured. The application can improve the overall efficiency and reliability of space governance when facing dynamic scenes or emergency situations.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of space governance, in particular to a multi-agent space collaborative governance method and system. BACKGROUND

[0002] In the field of space governance, satellite images, sensor real-time readings, policy texts and geographic coordinates and other types of information are needed to carry out work, and these information sources are scattered and have great differences in format, which are the core basis for supporting governance decision-making and agent scheduling. When dealing with such multi-source heterogeneous information, the current technology often lacks effective correlation means, making it difficult to establish semantic connections between different types of information, resulting in low information integration efficiency, and the generated governance foundation data is difficult to meet the unified and standardized application requirements.

[0003] At the same time, the environmental state and governance needs in the space governance scene are prone to dynamic changes, for example, in the event of sudden disasters or abnormal regional environmental parameters, agents need to quickly adapt to scene adjustments roles and tasks, but in the existing method, when facing dynamic scenes or emergency situations, response lag is prone to occur, affecting the overall efficiency and reliability of space governance.

[0004] From the above, how to improve the overall efficiency and reliability of space governance when facing dynamic scenes or emergency situations still needs to be solved. SUMMARY

[0005] In order to improve the overall efficiency and reliability of space governance when facing dynamic scenes or emergency situations, the application provides a multi-agent space collaborative governance method and system.

[0006] In a first aspect, the application provides a multi-agent space collaborative governance method, which adopts the following technical solution:

[0007] A multi-agent space collaborative governance method, comprising:

[0008] The multi-source heterogeneous data composed of satellite images, sensor real-time readings, policy texts and geographic coordinate data are processed, the corresponding cross-type semantic association is established based on geographic spatial data embedding technology, the satellite images are optimized by vLLM dynamic memory management, the regional environmental parameters and governance needs are associated to generate space attribute semantic expression, a unified structured text is generated, and the distributed consistency synchronization mechanism is used to transmit the text to the central database;

[0009] Based on the spatial environment information in the structured text, the governance target constraint and the agent capability characteristics, when the environment changes or the demand adjusts, the central server large language model dynamically generates an agent role image including the name, skill, knowledge boundary and spatial working range based on the preset Prompt template, and generates an image that does not depend on fixed rules, wherein the knowledge boundary includes policy and regulation adaptation instructions, and the spatial working range includes associated geographic coordinates.

[0010] The total task information of the role image and the spatial governance is obtained, and in the edge-center collaborative architecture, the target computing node is selected based on the total task characteristics combined with the semantic attributes of the role; the target node splits the total task information into subtasks with dependency relationships by using a Divide-and-Conquer algorithm, establishes semantic mapping between the subtasks and the role image, calculates the matching probability, and optimizes the distribution rules according to the scene to obtain the corresponding splitting and distribution results.

[0011] In the perception layer-decision layer-execution layer architecture, a corresponding multi-dimensional governance scheme is generated based on the splitting and distribution results combined with real-time spatial data, the agent task set is determined and the governance operation instruction is output, wherein the perception layer transmits the real-time data to the decision layer after being processed by the edge node, the decision layer integrates the agent suggestions based on the large language model and processes conflicts according to semantic priority, and the execution layer encodes geographic metadata and instructions based on the spatial perception protocol and transmits them to the corresponding agent.

[0012] The execution feedback information corresponding to the agent is obtained, and the role generation strategy and the distribution rules of the large language model are updated based on the execution feedback information; a multi-level trigger mechanism is established, wherein the sensor data threshold value starting condition trigger and the sudden disaster monitoring event trigger are automatically re-distributed roles.

[0013] Optionally, in the process of processing the multi-source heterogeneous information, the method further comprises:

[0014] The sensor real-time readings are first aligned with time series records by using a Kalman filter algorithm for outlier removal;

[0015] Before the vLLM dynamic memory management optimization, a multi-band image fusion technology is used to enhance the spatial environment feature extraction accuracy of satellite images;

[0016] Based on the enhanced satellite images and the aligned sensor readings, semantic association is established, so that the error rate of the spatial environment information in the generated structured text is reduced.

[0017] Optionally, in the process of generating the agent role image, the method further comprises:

[0018] The preset prompt template further comprises a role capability evaluation index, wherein the role capability evaluation index comprises spatial governance task proficiency, policy and regulation matching degree, and cross-regional cooperation response speed;

[0019] After the central server large language model generates the role image, the role capability evaluation index is updated in real time based on the region environment change information fed back by the perception layer, and the skill weight and spatial work range boundary in the role image are dynamically adjusted.

[0020] Optionally, in the process of obtaining the corresponding disassembly and distribution result, the method further comprises:

[0021] The geographical coordinate associated region based on the subtask adds a subtask space-time constraint determination step;

[0022] If the geographical ranges of adjacent subtasks overlap and the execution times conflict, the subtask with high governance demand urgency is retained preferentially, and the execution time sequence of another subtask is adjusted;

[0023] In the calculation of the matching probability, in addition to the comprehensive subtask query, the agent capability and the historical distribution result, the agent real-time task load rate is introduced, so that the matching probability calculation result is more consistent with the actual bearing capacity of the agent.

[0024] Optionally, the method further comprises:

[0025] Before integrating the agent suggestion, the decision layer calls a pre-built spatial governance scene knowledge base, which comprises historical governance cases and corresponding scheme effect records;

[0026] The real-time spatial information is compared with the historical information of similar scenes in the knowledge base through the large language model to generate a corresponding scheme feasibility evaluation report;

[0027] When the pass rate of the scheme in the evaluation report is greater than or equal to 85%, a corresponding final governance scheme is generated based on the agent suggestion, and before the execution layer outputs the governance operation instruction, the geographical metadata in the instruction is subjected to coordinate verification.

[0028] Optionally, in the process of automatically redistributing the roles, the method further comprises:

[0029] The feedback information is classified according to spatial governance types and region complexity, the spatial governance types include ecological monitoring, disaster emergency, and urban control, and the region complexity includes core urban area and remote mountainous area; wherein, a preset scene adaptation weight initial matrix is used to replace the fixed dimension weighting method;

[0030] The large language model performs semantic reflection on the feedback information, obtains the corresponding reflection result, stores the reflection result as situational memory and injects it into the subsequent Prompt, dynamically corrects the weight matrix based on the situational memory; if the feedback information shows that the role adaptation degree is lower than the threshold value for three times in succession, the tool evolution process is automatically started, the large language model generates a tool calling strategy based on the governance demand, and the skill boundary of the role image is updated at the same time.

[0031] Optionally, in addition to conditional triggering and event triggering, a capability decay triggering is added, and the method further comprises:

[0032] Based on the situational memory, if the success rate of the intelligent agent decreases by more than 30% with each task round, it is determined that the capability decays, and the role reassignment is automatically triggered; and the resource threshold triggering adopts a dynamic calculation model to adjust in real time according to the current governance task urgency;

[0033] When the standby agent is started in an extreme scenario, the historical optimal role-task matching track in the situational memory is migrated to the standby agent, and the role parameters are fine-tuned in combination with the real-time spatial information of the perception layer; at the same time, a lightweight model snapshot of the core role is pre-stored in the edge node, and the standby agent can directly call the snapshot to complete initialization when it is started, shortening the switching delay to seconds;

[0034] After the standby agent is running, the capability of the standby agent is verified through a spatial perception protocol with other node intelligent agents to ensure that the skill covers the current unfinished sub-tasks.

[0035] In a second aspect, the present application provides a multi-agent spatial collaborative governance system, which adopts the following technical solution:

[0036] A multi-agent spatial collaborative governance system comprises:

[0037] A multi-source heterogeneous data processing and storage module processes multi-source heterogeneous data composed of satellite images, real-time sensor readings, policy texts and geographic coordinate data, establishes corresponding cross-type semantic association based on geographic spatial data embedding technology, optimizes satellite images through vLLM dynamic memory management, associates regional environment parameters with governance demand to generate spatial attribute semantic expression, generates a unified structured text, and transmits the text to a central database through a distributed consistency synchronization mechanism;

[0038] A role image dynamic generation module dynamically generates an intelligent agent role image including name, skill, knowledge boundary and spatial working range based on spatial environment information, governance target constraints and agent capability characteristics in the structured text, generates an agent role image including name, skill, knowledge boundary and spatial working range based on a preset Prompt template by a central server large language model when the environment changes or the demand is adjusted, and generates an agent role image not dependent on fixed rules, wherein the knowledge boundary includes policy and regulation adaptation instructions, and the spatial working range includes associated geographic coordinates.

[0039] The total task decomposition intelligent allocation module obtains total task information of the role image and space governance, selects a target computing node based on total task characteristics combined with role semantic attributes in an edge-center collaborative architecture; the target node splits the total task information into subtasks containing dependency relationships using a Divide-and-Conquer algorithm, establishes semantic mapping of the subtasks and the role image, calculates a matching probability, and obtains corresponding decomposition and allocation results by constructing a reward function according to a scene to optimize allocation rules.

[0040] The governance scheme generation and instruction output module generates corresponding multi-dimensional governance schemes based on the decomposition and allocation results combined with real-time space data in a perception layer-decision layer-execution layer architecture, determines an agent task set and outputs governance operation instructions, wherein the perception layer transmits real-time data to the decision layer after being processed by an edge node, the decision layer integrates agent suggestions based on a large language model and processes conflicts according to semantic priorities, and the execution layer encodes geographic metadata and instructions based on a space perception protocol and transmits them to corresponding agents.

[0041] The feedback optimization and multi-level trigger guarantee module obtains execution feedback information of the agents, updates role generation strategies and allocation rules of the large language model based on the execution feedback information, and sets up a multi-level trigger mechanism, wherein a sensor data threshold value starting condition trigger and a sudden disaster starting event trigger are both automatically re-allocated roles.

[0042] In a third aspect, the present application provides a multi-agent space collaborative governance system, which adopts the following technical solution:

[0043] A multi-agent space collaborative governance system includes a processor, and the processor runs a program of the multi-agent space collaborative governance method of any one of the above.

[0044] In a fourth aspect, the present application provides a storage medium, which adopts the following technical solution:

[0045] A storage medium stores a program of the multi-agent space collaborative governance method of any one of the above.

[0046] In summary, the present application includes at least one of the following beneficial technical effects:

[0047] In a dynamic scene, the multi-source heterogeneous data processing link is subjected to outlier rejection, image feature enhancement, and vLLM dynamic memory management, which not only ensures the accuracy of spatial environment information but also speeds up the data processing speed, and in combination with a distributed synchronization mechanism, real-time data support is provided for subsequent decision-making; the agent role portrait can be based on regional environmental changes to update the ability index and working range at regular intervals, and when task allocation, time and space constraints are used to avoid task conflicts, real-time load rate is introduced to optimize matching accuracy, and the decision layer quickly generates a highly feasible solution with the help of a scene knowledge base, and the dynamic adaptation capability of each link greatly reduces resource waste and decision delay, realizing efficient connection of the governance process.

[0048] In the face of emergency situations, a multi-level triggering mechanism (sensor threshold, sudden disaster, agent capability degradation) can capture abnormalities in real time and automatically start role redistribution to avoid governance gaps; in extreme scenarios, backup agents rely on historical optimal matching trajectories and lightweight model snapshots to achieve second-level startup and parameter fine-tuning, while cross-node capability collaboration verification ensures skill coverage of incomplete tasks, combined with feedback-driven agent tool evolution, effectively addressing potential capability gaps during the emergency process, ensuring the continuous and stable operation of the governance system in the event of an emergency, and significantly reducing the risk of governance failure due to system interruption or insufficient capacity. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 is a flowchart of a multi-agent spatial collaborative governance method according to an exemplary embodiment.

[0050] Figure 2 is a structural block diagram of a multi-agent spatial collaborative governance system according to an exemplary embodiment. DETAILED DESCRIPTION

[0051] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0052] In the description of the present specification, the description of the terms "certain embodiments", "one embodiment", "some embodiments", "illustrative embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the described embodiments or examples are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0053] The embodiments of the present application disclose a multi-agent spatial collaborative governance method, referring to Figure 1 , comprising:

[0054] S100, processing multi-source heterogeneous data composed of satellite images, sensor real-time readings, policy texts, and geographic coordinate data, establishing corresponding cross-type semantic associations based on geospatial data embedding technology, optimizing satellite images through vLLM dynamic memory management, associating regional environmental parameters and governance needs with geographic coordinates to generate spatial attribute semantic expressions, generating unified structured text, and transmitting it to the central database through a distributed consistency synchronization mechanism.

[0055] S100 specifically includes the following steps:

[0056] Step 1: Classification and initial arrangement of multi-source heterogeneous data:

[0057] In the initial stage of S100, first, classify and sort the four types of core multi-source heterogeneous data: satellite images, sensor real-time readings, policy texts, and geographic coordinate data. Due to significant differences in format, dimension, and information density among different types of data, such as high-dimensional visual data for satellite images, time-series numerical data for sensor readings, unstructured text data for policy texts, and spatial location data for geographic coordinates, independent preprocessing channels need to be established according to data types to eliminate obvious invalid information (such as blank readings when sensors are offline, extreme noise regions in satellite images), laying a foundation for subsequent targeted processing and avoiding processing bias caused by data mixing.

[0058] Step 2: vLLM dynamic memory management optimization for satellite images:

[0059] To address the problem of large satellite image data volume and low processing efficiency, vLLM dynamic memory management technology is used for optimization. This technology dynamically adjusts the memory allocation mechanism of large language models (LLM) when loading satellite image data, and adaptively allocates memory resources based on image resolution, band number, and regional coverage: for high-resolution, multi-band key area images, priority is given to memory supply to ensure that detailed information is not lost; for low-priority background area images, memory reuse mechanisms are used to reduce resource occupation. At the same time, the pre-compilation and batch processing capabilities of vLLM can improve the feature extraction speed of satellite images, avoiding delays in image processing that affect the timeliness of the overall governance process.

[0060] Step 3: Geographic coordinate association and spatial attribute semantic expression generation:

[0061] With geographic coordinate data as the core link, it is deeply associated with regional environmental parameters and governance needs to generate spatial attribute semantic expression. Specifically, first, through geocoding technology, discrete geographic coordinates are mapped to actual governance areas (such as streets, watersheds, and ecological protection areas), and the specific spatial range corresponding to each coordinate point is determined. Then, the environmental parameters of the area (such as air quality index, water quality index, and vegetation coverage) are retrieved, combined with the current stage of governance needs (such as pollution control, ecological restoration, and disaster prevention), and the association of "coordinate position + environmental state + governance target" is converted into structured semantic description. For a certain river basin, the current water quality COD value is 45 mg / L, and the governance demand is to reduce the COD value to below 30 mg / L, so that the originally single coordinate data has practical significance in the governance scenario.

[0062] Step 4, cross-type semantic association establishment (based on geographic spatial data embedding technology):

[0063] With the help of geographic spatial data embedding technology, the semantic barriers between different types of data are broken down, and a unified cross-type semantic association is established. This technology maps various types of data to the same geographic spatial semantic dimension: for satellite images, extract their pixel-level spatial features and convert them into semantic vectors (such as "image feature vector of vegetation coverage area in a certain region"); for sensor real-time readings, combine the geographic coordinates of their collection locations to generate "location + time series value" semantic vectors (such as "PM2.5 value vector of sensor at 30° N XX' for 24 hours"); for policy texts, extract keywords related to spatial governance (such as "watershed non-discharge range" and "ecological red line area") and associate corresponding geographic coordinates to generate "policy provisions + spatial range" semantic vectors. Through vector similarity calculation, different types of data form semantic associations around the same geographic spatial scene, such as automatically associating "abnormal water area" in satellite images with "water quality exceeding standard readings" of sensors and "water governance provisions" in policy texts, realizing the fusion of data information.

[0064] Step 5, unified structured text generation and distributed synchronous storage:

[0065] Based on the above processing results, a unified format structured text is generated. The structured text needs to include four core fields: "data source type, geographic spatial range, core information content, and semantic association identifier", for example, "data source: satellite image; geographic spatial range: 30° N XX' N, 120° E XX' E; core information content: abnormal water area; semantic association identifier: water quality exceeding standard readings, water governance provisions". - , 120° E - Core information: There are 2 vegetation wilting areas in this region; Semantic association identifier: Associated sensor reading (No. S001) shows that soil humidity has been below the threshold for nearly 7 days, and associated policy text (No. P005) is 'Vegetation Protection and Repair Regulations'. After completing the structured text generation, the text is transmitted to the central database through a distributed consistent synchronization mechanism: This mechanism ensures the accuracy of text transmission through multi-node data verification (edge nodes and central nodes are checked bidirectionally), while using an incremental synchronization strategy (only transmitting updated or new text data) to reduce bandwidth occupation, ensuring that the central database can obtain unified multi-source data processing results in real time and reliably, avoiding subsequent decision biases caused by asynchronous data.

[0066] Through classification and targeted optimization (such as satellite image vLLM memory management), the efficiency and accuracy of data processing are improved; through geographical coordinate association and cross-type semantic association, data barriers are broken down, and information fusion is achieved; through unified structured text and distributed synchronization, a solid data foundation is laid for subsequent S200 intelligent agent role profiling (providing accurate spatial environment information and governance demand data), S300 task decomposition and allocation (providing unified data semantic logic), and S400 governance scheme generation (providing real-time spatial data support), ensuring the accuracy, real-time nature, and collaboration of the entire spatial governance process from the source, avoiding governance decision-making errors or inefficiencies caused by data problems.

[0067] S200, based on the spatial environment information, governance target constraints, and agent capability characteristics in the structured text, generates an intelligent agent role profile including name, skills, knowledge boundaries, and spatial working range based on a pre-set Prompt template from the central server large language model when the environment changes or demand adjusts. The generation does not rely on fixed rules, and the knowledge boundaries include policy and regulation adaptation instructions, and the spatial working range includes associated geographical coordinates.

[0068] Specifically, S200 includes the following steps:

[0069] Step 1: Extraction and integration of input information

[0070] In the initial stage of S200, first extract the core information related to role profile generation from the unified structured text generated by S100, including spatial environment information, governance target constraints, and basic capability characteristics of the agent to form a complete input data set.

[0071] Among them, the spatial environment information is obtained from the "geospatial range" and "core information content" fields of the structured text, such as the vegetation coverage state, water quality condition, etc. of a certain area; the governance target constraint corresponds to the associated governance demand in the structured text, such as "improve the water quality of a certain watershed to the standard level" and "control the dust pollution in a certain area"; and the agent basic capability feature includes the existing hardware support (such as data processing rate, communication coverage) and existing functional modules (such as environment monitoring module, instruction execution module) of the existing agent. Through the classification extraction and cross verification of the three types of information, it is ensured that the input data can accurately reflect the actual demand of the current governance scene and the basic capability of the agent, and the disconnection between the profile and the scene caused by information missing is avoided.

[0072] Step 2, determination of trigger condition and model calling:

[0073] Based on the extracted input information, the two types of dynamically generated trigger conditions, "environment change" and "demand adjustment", are determined in real time, and then the central server large language model is started. For the determination of "environment change", by comparing the current extracted spatial environment information with the historical same period or the structured text information of the previous period, if the change amplitude of the key environmental indicators (such as the PM2.5 concentration in a certain area, the dissolved oxygen content of water body) exceeds the preset threshold, it is determined that the environment has changed; for the determination of "demand adjustment", according to the update of the governance target constraint, such as the original target is "short-term disaster emergency monitoring" and the current adjustment is "long-term ecological restoration tracking", the generation process is triggered. When any trigger condition is met, the central server large language model is automatically called, and a real-time data interaction channel between the model and the input data set is established to provide computing support for subsequent profile generation.

[0074] Step 3, parameter injection of preset Prompt template:

[0075] After calling the large language model, the preset Prompt template is used as the generation framework, and the input information extracted in step 1 is injected into the template structure according to the template structure. The preset Prompt template includes four core modules: "scene positioning", "target matching", "capability adaptation" and "space binding". In the "scene positioning" module, the spatial environment information is injected to clearly define the governance area and environmental background corresponding to the profile; in the "target matching" module, the governance target constraint is injected to define the specific governance task supported by the profile; in the "capability adaptation" module, the agent basic capability feature is injected to ensure that the generated skills are compatible with the agent hardware and existing modules; and in the "space binding" module, a geospatial coordinate association interface is reserved to prepare for subsequent spatial work range generation. Through parameter injection, the large language model can generate a profile around a specific scene, avoiding non-targeted generalization output, and ensuring the compatibility of the profile with the actual governance demand and the agent capability.

[0076] Step 4: Dynamic generation of role image core elements:

[0077] The large language model generates the four core elements of the role image based on the Prompt template with injected parameters, and the generation process does not rely on fixed rules. Among them, "name" is generated by combining scene positioning and target matching results, which needs to reflect the governance function and regional association, such as "certain river basin water quality monitoring agent" and "certain urban dust control agent"; "skills" are dynamically matched according to governance target constraints and agent capability characteristics, such as "water quality index real-time analysis" and "governance measure effect feedback" for the "water quality improvement" target, rather than using a fixed skill list; "knowledge boundary" contains policy and regulation adaptation instructions, which clearly indicate the policy guidance and operation standards that the agent needs to follow when carrying out governance in the corresponding region, such as "adapt to regional water resources protection related policy requirements and clearly define the compliance operation standards for water intake and drainage links"; "spatial working range" is generated by associating the geographical coordinates in the structured text, which accurately defines the governance coverage area of the agent, such as "certain river basin range with latitude - , longitude - , to ensure that the agent only performs tasks within the specified space and avoids cross-regional resource waste.

[0078] Step 5: Verification and associated storage of role image:

[0079] After the role image is generated, it needs to be verified in multiple dimensions to ensure the completeness and adaptability of the elements. The verification checks whether the "skills" cover the core needs of the governance target, whether the "knowledge boundary" is consistent with the regional policy guidance, and whether the "spatial working range" is accurately matched with the geographical coordinates. If there is a deviation, the large language model will be returned to adjust. After verification, the role image is associated with the structured text of S100 (indexed by geographical coordinates) and stored in the image dedicated module of the central database. When storing, the image is added with a dynamic update identifier, recording the generation time and trigger condition type, which facilitates subsequent monitoring of new environmental changes or demand adjustments, and quickly retrieves historical images for iterative optimization. At the same time, it ensures that S300 can directly associate the corresponding spatial environment information and governance target when obtaining the role image, reducing data query time consumption.

[0080] By dynamic trigger condition judgment, the image can be updated in real time according to the environment and demand changes, avoiding the "disconnection of ability" of fixed image after the scene changes; through the Prompt template parameter injection and core element dynamic generation, it not only ensures the high adaptation of the image and the governance demand, the ability of the agent, but also improves the scene adaptability through the characteristics of "not relying on fixed rules", especially in complex spatial governance scenarios, it can generate more targeted role functions; through verification and associated storage, it ensures that the image quality is reliable and easy to call later. Overall, S200 as the core link connecting "data processing" and "task allocation" provides accurate "agent ability template" for the total task decomposition and intelligent allocation of S300, and through the binding of spatial working range and geographic coordinates, the adaptation of knowledge boundary and policy guidance, the task execution of the agent in the subsequent governance process is more spatially accurate and compliant, which ensures the efficiency and reliability of spatial governance from the aspect of "agent adaptability", avoiding the waste of governance resources or task execution deviation caused by the ambiguity of agent role and the mismatch of ability.

[0081] S300, obtaining the total task information of the role image and spatial governance, in the edge-center collaborative architecture, selecting the target computing node based on the total task characteristics combined with the semantic attributes of the role; the target node uses the Divide-and-Conquer algorithm to split the total task information into subtasks with dependency relationship, establishes the semantic mapping of subtasks and role image, calculates the matching probability, and optimizes the allocation rules according to the scene to obtain the corresponding decomposition and allocation results.

[0082] Among them, S300 specifically includes the following steps:

[0083] Step 1, integration and acquisition of role image and total task information:

[0084] First, the agent role image set generated by S200 is retrieved from the center database, and the total task information issued by the spatial governance system is received. The role image needs to include the skill type, knowledge boundary, spatial working range (associated with geographic coordinates), and real-time load state of the agent, and the total task information needs to clearly define the governance target (such as ecological monitoring in a certain area, environmental remediation after disaster), spatial coverage, execution time limit, and core requirements (such as data accuracy, response speed). Preliminary verification is performed on the two types of information to ensure that the role image has no information missing and the total task parameters have no logical contradictions (such as avoiding the total task spatial range exceeding the working range of all roles), providing complete and accurate basic information for subsequent node selection and task decomposition, avoiding process jam caused by incomplete information.

[0085] Step 2, target computing node selection in edge-center collaborative architecture:

[0086] Based on the resource distribution characteristics of the edge-center collaborative architecture, the target computing node is determined in combination with the total task characteristics and the role semantic attributes. The total task characteristics focus on three aspects: first, the urgency (such as high urgency for disaster management, and regular emergency for daily environmental monitoring), second, the spatial coverage range (such as cross-city regional tasks that require central node coordination, and single street tasks that can be handled by edge nodes), and third, the calculation complexity (such as multi-regional data correlation analysis requiring high-power nodes, and simple data statistics can be handled by edge lightweight nodes); the role semantic attributes focus on skill matching degree (such as when the total task contains image recognition requirements, the role node with image analysis skills is preferentially associated), and spatial work range overlap degree (such as when the role work range and the total task area overlap more than 80%, the edge node where the role is located is preferentially considered). The finally selected node needs to meet the requirements of "total task processing capacity adaptation" and "role coordination convenience", for example, high-urgency cross-regional tasks select central nodes, and regular local tasks select edge nodes, to achieve efficient use of architecture resources.

[0087] Step 3, Divide-and-Conquer algorithm to realize total task decomposition:

[0088] The target computing node calls the Divide-and-Conquer algorithm to split the total task into subtasks with dependency relationships according to the logic of "first large module splitting, then sub-module refinement". First, the total task is decomposed into several independent core modules, for example, "certain watershed management total task" is split into "watershed water quality monitoring", "pollution source positioning", and "governance measures implementation" modules; then, for each core module, the spatial work range of the role is further split into subtasks, such as the "watershed water quality monitoring" module is split into "upstream monitoring subtask", "midstream monitoring subtask", and "downstream monitoring subtask" according to geographical coordinates. The key is to clarify the dependency relationship between subtasks: for example, the "pollution source positioning" subtask needs to be completed first before starting the "governance measures implementation" subtask; the "upstream monitoring subtask" is completed first, and then the midstream and downstream monitoring subtasks are simultaneously promoted to avoid governance failure due to task order confusion. At the same time, during the decomposition process, it is necessary to ensure that the spatial range of the subtasks does not overlap, the time nodes do not conflict, and the complexity of a single subtask is adapted to the processing capacity of the agent.

[0089] Step 4, semantic mapping of subtasks and role profiles and calculation of matching probability:

[0090] Based on the core requirements of sub-tasks and the semantic attributes of role profiles, a two-way semantic mapping relationship is established, for example, the requirement of the "upstream water quality monitoring sub-task" is "possessing water quality parameter analysis ability and covering the upstream of the river basin in the working range", which corresponds to the role with "water quality analysis skills" and the spatial working range containing "upstream coordinates of the river basin", forming a semantic association of "sub-task requirement-role capability". Then the matching probability of the two is calculated, and the calculation process integrates three factors: first, the sub-task query matching degree (the semantic overlap degree of sub-task requirements and role skills, such as the matching degree of "water quality monitoring" and "water quality analysis"), second, the agent capability adaptation degree (the adaptation of the current capability level of the role to the difficulty of the sub-task, such as the adaptation of a senior role to a high-difficulty monitoring task), and third, the historical allocation effect (the success rate of the role in processing similar sub-tasks in the past, such as a success rate of over 90% to increase the matching probability weight). Through multi-dimensional calculation, it is ensured that the matching probability can truly reflect the suitability of the role in processing the sub-task, avoiding the problem of "capability mismatch".

[0091] Step 5, scenario-based reward function construction and allocation rule optimization:

[0092] The reward function is constructed in combination with the characteristics of the current governance scene to optimize the final task allocation rule. The core parameters of the reward function are dynamically adjusted according to the scene: for example, in the emergency governance scene (such as environmental assessment after an earthquake), the "task completion timeliness" weight accounts for 40%, the "result accuracy" weight accounts for 30%, and the "resource consumption" weight accounts for 30%; in the daily governance scene, the "result accuracy" weight accounts for 40%, the "resource consumption" weight accounts for 30%, and the "timeliness" weight accounts for 30%. The function sets reward items and penalty items: when the role matching probability is high and meets the scene weight demand (such as a role with high matching probability and fast response speed in an emergency scene), a high reward value is given; when the role matching probability is low or does not meet the scene demand (such as a role with high matching probability but high resource consumption in a daily scene), a penalty value is given. Based on the calculation results of the reward function, the allocation rule is adjusted, and the sub-tasks are preferentially allocated to the roles with the highest reward value, while ensuring that no role undertakes sub-tasks beyond its own load, and finally the explicit "sub-task-role" corresponding relationship is output, i.e. the disassembly and allocation results.

[0093] The edge-center collaborative selection and divide-and-conquer algorithm disassembles, which not only adapts to the processing needs of different types of tasks, but also ensures the orderly progress of tasks through the dependence relationship, avoiding resource waste and process confusion; the combination of semantic mapping and scenario-based reward function greatly improves the matching accuracy of sub-tasks and roles, reducing the efficiency loss caused by "capability mismatch". At the same time, the disassembly and allocation results output by this step directly provide "sub-task list" and "execution role positioning" for S400 governance scheme generation, ensuring that the subsequent scheme can accurately correspond to specific tasks and agents, from the task allocation level to guarantee the efficiency and orderliness of spatial governance, avoiding governance delay or failure caused by unreasonable task disassembly and improper role allocation.

[0094] S400, in the perception layer-decision layer-execution layer architecture, based on the disassembly and allocation results, generates corresponding multi-dimensional governance schemes combined with real-time spatial data, determines the agent task set and outputs the governance operation instructions, wherein the perception layer transmits the real-time data to the decision layer after being processed by the edge node, the decision layer integrates the agent suggestions based on the large language model and processes conflicts according to semantic priority, and the execution layer encodes geographic metadata and instructions based on the spatial perception protocol and transmits them to the corresponding agent.

[0095] S400 specifically includes the following steps:

[0096] Step 1, real-time data processing and transmission of the perception layer

[0097] In the initial stage of S400, the perception layer first completes the collection and preprocessing of real-time spatial data, and performs localized processing through the edge node. The real-time data collected by the perception layer covers multiple types of dynamic information, such as real-time updated regional environmental indicators (such as soil moisture, air quality values) by sensors, the latest satellite images, and spatial state changes (such as the density of people flow in a certain area, traffic flow) recorded by mobile monitoring devices.

[0098] The edge node processes these data in a targeted manner: denoising and time series alignment are performed on time series data (such as continuous sensor readings) to avoid misjudgment caused by data fluctuations; local feature extraction is performed on image data (such as satellite updated images) to focus on areas related to governance tasks (such as ecological restoration areas, disaster hazard areas); at the same time, all processed data are associated with corresponding geographic coordinates to mark the specific spatial range to which the data belong (such as a street, a watershed).

[0099] After processing, the edge node transmits standardized real-time spatial data to the decision layer through a low-latency transmission link, ensuring that the decision layer can obtain the latest and most accurate spatial state information to provide dynamic data support for governance scheme generation.

[0100] Step 2, multi-dimensional governance scheme generation and task set determination of the decision layer:

[0101] After receiving the real-time spatial data transmitted by the perception layer, the decision layer combines the task decomposition and allocation results output by S300 to start the generation process of the multi-dimensional governance scheme. First, the decision layer calls the large language model, taking "real-time spatial data + sub-task allocation list" as input, and synchronously collects governance suggestions proposed by each agent based on its role image (generated by S200), such as the agent responsible for ecological restoration may suggest "supplement irrigation for areas with soil humidity below the threshold", and the agent responsible for traffic control may suggest "optimize the route in areas with high human flow density".

[0102] The large language model integrates these suggestions at the semantic level: first, it identifies the matching degree of the suggestions with sub-tasks and real-time data (such as whether the irrigation suggestion corresponds to the area with abnormal soil humidity data), and then handles suggestion conflicts according to semantic priority, such as when "disaster emergency evacuation suggestions" conflict with "routine traffic control suggestions", the emergency class suggestions are prioritized; when there are differences in the same type of suggestions (such as different suggestions on irrigation volume from different agents), the optimal scheme is determined based on historical governance effect data and real-time environmental parameters (such as the current water retention capacity of the soil).

[0103] Finally, the decision layer generates a multi-dimensional governance scheme covering "spatial range, governance target, implementation steps, and responsible agents", and based on the scheme, it determines the specific task set for each agent (such as "agent A is responsible for irrigation operations in XX area, and agent B is responsible for human flow evacuation in XX road segment").

[0104] Step 3, instruction encoding and precise transmission at the execution layer:

[0105] After the decision layer determines the agent task set, the execution layer completes the encoding and transmission of the governance operation instructions. The execution layer first standardizes the encoding of geographic metadata and operation instructions in the governance scheme based on the spatial perception protocol: geographic metadata encoding needs to clearly specify the precise spatial information corresponding to the task, including the latitude and longitude range, boundary coordinates, and key point identifiers (such as the valve position of the irrigation area and the entrance and exit of the evacuation area) of the target area; operation instruction encoding needs to convert abstract tasks into specific action instructions executable by agents, such as encoding "supplement irrigation" into "start XX type irrigation equipment, set flow XX cubic meters / hour, duration XX minutes", and encoding "human flow evacuation" into "play guiding voice at XX point, update human flow density broadcast every 10 minutes".

[0106] After encoding is completed, the execution layer transmits the encoded instructions to the corresponding agent through the exclusive communication link established by the spatial awareness protocol. This link can optimize the transmission path according to the spatial position (associated geographic coordinates) of the agent, avoiding instruction transmission delay or misfire. At the same time, a check code is added during the instruction transmission process to ensure that the instructions received by the agent are complete and tamper-free, ensuring the consistency of subsequent execution actions and governance solutions.

[0107] The edge node processing of the perception layer ensures the effectiveness of real-time spatial data, providing dynamic basis for solution generation and avoiding solutions that deviate from reality due to data lag. The large language model integration and semantic priority conflict processing of the decision layer not only fully utilizes the professional capabilities of each agent but also ensures the uniformity and emergency adaptability of the solution, avoiding governance direction deviation caused by chaotic suggestions. The standardized encoding and precise transmission of the execution layer ensure that governance instructions can be accurately and efficiently transmitted to the corresponding agent, avoiding execution errors caused by ambiguous or incorrect instructions. Overall, S400 takes over the task allocation results of S300, connects the actual execution actions of the agent, and through the division and cooperation of the three-layer architecture, significantly improves the scientificity, pertinence, and landing efficiency of the spatial governance solution, providing a key guarantee for the subsequent precise execution of the agent's governance tasks.

[0108] S500, obtain execution feedback information corresponding to the agent, and update the role generation strategy and allocation rules of the large language model based on the execution feedback information; set a multi-level triggering mechanism, wherein the sensor data threshold value start condition trigger and the sudden disaster start event trigger are automatically re-allocated roles.

[0109] S500 specifically includes the following steps:

[0110] Step 1, multi-dimensional collection and classification of agent execution feedback information:

[0111] In the initial stage of S500, the execution feedback information of each agent is first collected and classified. The feedback information covers the core state data of the agent's task execution, such as task completion progress (e.g., "85% of the regional environmental monitoring task is completed"), resource consumption in the execution process (e.g., CPU occupancy, data transmission delay), governance effect matching degree (e.g., "the actual water quality improvement value deviates from the target value by 10%"), and abnormal situations encountered in the execution process (e.g., "the sensor is temporarily offline, causing data collection interruption"). Due to the significant differences in the use of different types of feedback information (e.g., completion progress for evaluating task effectiveness, resource consumption for optimizing allocation rules), a classified storage directory is established according to "task effectiveness class, resource consumption class, and abnormal event class", and the corresponding agent ID and task number are associated to ensure that each feedback information can be traced back to a specific role and task, laying a foundation for subsequent accurate analysis and avoiding strategy update deviation caused by chaotic feedback information.

[0112] Step 2, semantic analysis and key feature extraction of feedback information:

[0113] For the collected classified feedback information, a large language model is used for semantic analysis and key feature extraction to convert unstructured feedback descriptions into quantifiable and correlatable feature data. For example, for the text feedback "a certain agent's skill deficiency leads to substandard governance effect", the model will analyze the core problem of "low role adaptation" and extract key features such as "involved agent skill type, substandard governance indicator, and deviation value". For the abnormal feedback "a certain regional sensor reading exceeds the threshold for 5 consecutive minutes", the model will extract features such as "abnormal sensor location, threshold parameter type, and threshold exceeding duration". At the same time, through semantic association technology, the feedback features are bound with the role image attributes generated by S200 (such as skills, spatial working range), and the task allocation results of S300 (such as subtask type, matching probability), to clarify the association between feedback problems and previous role generation and task allocation links, and provide accurate direction basis for strategy update.

[0114] Step 3, updating the role generation strategy of the large language model based on feedback features:

[0115] In combination with the extracted feedback key features, the role generation strategy of the large language model is dynamically updated, and the generation logic of the subsequent role image is optimized. For example, if multiple feedback information shows that "the spatial working range of certain agents is too wide, causing response delay", the parameter of "spatial working range division" in the role generation Prompt template is adjusted, and the original rule of dividing by large area is optimized to the fine division of "core governance point + radiation range"; if the feedback shows that "the processing capacity of some agents for certain environmental parameters is insufficient", the processing capacity weight of the corresponding parameters is added in the role skill generation link to ensure that the subsequent generated role image is more in line with the actual governance needs. In the updating process, it does not rely on fixed rules, but relies on the statistical law of feedback features, such as when the proportion of "skill deficiency" feedback exceeds 30%, the iteration of skill generation logic is automatically triggered, avoiding the disconnection between roles and needs caused by fixed strategies.

[0116] Step 4, optimizing task allocation rules based on feedback features:

[0117] Based on the resource consumption and task effectiveness features in the feedback information, the task allocation rules in S300 are optimized to improve the rationality and efficiency of subsequent task allocation. For example, if the feedback shows that "a certain computing node has high load and execution delay due to the allocation of too many sub-tasks", the selection logic of the target computing node is adjusted, and "node historical load rate" is added as a selection basis to preferentially select nodes with load below the threshold; if the feedback shows that "although the matching probability of sub-tasks and roles is high, the actual load of the role is not considered, resulting in poor effectiveness", a "role real-time load coefficient" is added when calculating the matching probability to modify the original matching probability, making the allocation result more in line with the actual carrying capacity of the agent. At the same time, in combination with the task dependency relationship problems in the feedback (such as "predecessor sub-task delay causes subsequent task stall"), the decomposition logic of the Divide-and-Conquer algorithm is optimized, and "sub-task time sequence priority" is added to ensure more reasonable dependency relationships.

[0118] Step 5, deployment of multi-level trigger mechanism and automatic role reallocation execution:

[0119] A multi-level mechanism of "conditional triggering + event triggering" is deployed, and automatic role reallocation is executed when the trigger condition is met to ensure the continuity of the governance process. For conditional triggering: real-time monitoring of environmental parameter data transmitted by sensors (such as air quality, water quality, temperature), when a certain parameter exceeds the preset threshold (such as PM2.5 concentration exceeding 150 When the abnormal parameter condition is triggered, the role reassignment is automatically triggered, the updated role generation strategy is invoked, the temporary role that adapts to the abnormal parameter condition is quickly generated, and the related subtasks are re-assigned to the new role based on the optimized assignment rule S300; for event triggering: through the real-time monitoring module of the edge node (such as a disaster monitoring camera, a vibration sensor), a sudden disaster (such as a flood, a landslide) is captured, and once the disaster signal is monitored, the emergency reassignment process is triggered immediately, the non-critical tasks are suspended, the governance role resources in the disaster area are prioritized, the role with disaster emergency processing skills is invoked, and the subtasks in the disaster area are quickly taken over, while the task assignment permission of the faulty agent in the disaster area is closed synchronously to avoid invalid execution.

[0120] Through the classified collection and accurate analysis of feedback information, the deficiencies in the early role generation and task assignment links are clarified to provide data support for strategy optimization; based on the feedback characteristics, the role generation strategy and the assignment rule are updated to realize the closed-loop iteration of "governance-feedback-optimization" and continuously improve the adaptation of the role and the task; the deployment of the multi-level triggering mechanism can quickly respond to environmental abnormalities or sudden disasters and ensure the continuity of the governance process through automatic role reassignment.

[0121] Overall, S500 not only solves the possible static problem of the early process, but also provides emergency protection capability for the governance system, ensuring the long-term effectiveness and stability of the multi-agent spatial collaborative governance from the two dimensions of "continuous optimization" and "risk response".

[0122] Taking the emergency scenario of "urban watershed sudden rainstorm causing water pollution" as an example, the scheme first responds quickly through the multi-level triggering mechanism of S500: 1 hour after the rainstorm, the water quality sensor in the watershed monitors that the COD value exceeds 120 mg / L (threshold value is 80 mg / L) for 3 consecutive times, and the condition triggering mechanism is automatically triggered; at the same time, the disaster monitoring camera of the edge node captures the garbage along the riverbank flowing into the water, and the event triggering is started. At this time, S100 immediately processes the multi-source data, the satellite image is optimized by vLLM dynamic memory management, the pollution diffusion range of the watershed is quickly extracted, the real-time readings of the sensor are aligned by Kalman filtering, and the pollution concentration gradient is clarified; the geographical coordinates are associated with the watershed partition (upper stream, middle stream urban section, lower stream wetland) and the governance demand of "controlling pollution diffusion within 24 hours", the structured text is generated and synchronized to the central database, and the data integration is completed only in 20 minutes, providing accurate and real-time data support for emergency governance and avoiding the expansion of the pollution range due to data lag.

[0123] Subsequently, the scheme realizes role adaptation and task efficient allocation through S200 and S300: the center server large language model dynamically generates three types of roles of "river basin water quality monitoring agent", "pollution source interception agent" and "emergency dredging agent" based on structured text, the spatial working range of the former is bound to the upstream tributary coordinates, and the skill focuses on COD real-time analysis; the latter two cover the middle reaches of the urban section and the lower reaches of the wetland, and the knowledge boundary contains emergency operation specifications. In the edge-center collaborative architecture, three edge nodes (handle local monitoring and interception tasks) and one center node (overall dredging scheduling) are selected around the selected river basin; the target node divides the "total pollution control task" into three types of sub-tasks of "pollution range real-time monitoring", "coastal interception facility start" and "key area dredging" by using the Divide-and-Conquer algorithm, matches the probability by combining the real-time load rate of the agent (such as assigning "high concentration pollution area monitoring" to the water quality monitoring agent with a load rate lower than 40%), and preferentially guarantees the emergency task by using the reward function (the reward weight of "interception start" is set to 50%), and the task allocation is completed within 15 minutes to avoid delay of governance due to role mismatch or task accumulation.

[0124] Finally, S400 is executed and S500 is fed back to optimize and guarantee the reliability of governance: the edge node of the perception layer transmits pollution concentration update data to the decision layer every 10 minutes, and the decision layer large language model integrates the agent suggestion, when the "water quality monitoring agent" feedback "midstream pollution still spreads" and the "pollution source interception agent" suggestion "add interception point" conflict, the latter is reserved according to the "emergency priority", and the governance scheme of "adding 2 temporary interception dams" is generated; the execution layer encodes "interception dam coordinates + start timing" into instructions and transmits them to the corresponding agent based on the spatial perception protocol. After the execution feedback shows that "the downstream dredging efficiency is low", S500 immediately updates the role generation strategy (adds the "large equipment cooperation" skill to the dredging agent) and optimizes the allocation rules (the next similar task is preferentially selected for the agent with a historical dredging efficiency of more than 90%); during this period, one edge node fails, triggering S500 role reallocation, calling the standby dredging agent (loading the historical optimal parameter snapshot, completing initialization within 3 seconds) to take over the task, and finally controlling the COD value of the river basin to be within the threshold within 48 hours, which is 3 times the efficiency of traditional manual governance, and there is no interruption of governance, which fully embodies the efficiency and reliability of the scheme in emergency scenarios.

[0125] In the embodiments of the present application, in the process of processing multi-source heterogeneous information, the method further comprises:

[0126] Step 1, after S100 "Multi-source heterogeneous data classification and initial arrangement" and before "Cross-type semantic association establishment", Kalman filtering algorithm is performed for sensor real-time readings (such as time series data of basin water quality pH value and turbidity): first, the "state prediction-observation update" cycle is constructed through the algorithm, the current time theoretical value is predicted based on the historical reading law of the sensor, and the predicted deviation is corrected combined with the actually collected readings to eliminate invalid data caused by sensor fluctuations (such as instantaneous abnormal readings caused by rainstorm impact); at the same time, the corrected readings are time-aligned according to the time stamp, ensuring that the data of different position sensors in the same time period can be compared horizontally, providing a regular time series data basis for subsequent correlation analysis.

[0127] Step 2, before S100 "vLLM dynamic memory management optimization of satellite images", multi-band image fusion technology is used for satellite images (such as visible light band and near-infrared band images of the basin): select key bands related to basin pollution monitoring (such as visible light band reflecting water turbidity and near-infrared band identifying pollutant types), superimpose the feature information of multi-band images through pixel-level fusion algorithm, retain the advantage features of each band (such as basin boundary details of visible light band and pollutant distribution characteristics of near-infrared band), weaken the noise interference of single band, and generate a fusion image with high feature recognition, laying a foundation for accurate extraction of environmental features during subsequent vLLM optimization processing.

[0128] Step 3, after the alignment of sensor readings and the enhancement of satellite images, the S100 "cross-type semantic association establishment" section is added with targeted association operations: using geographic coordinates as a link, the spatial features such as "basin pollution patch position and area" extracted from the enhanced satellite images are bound with the time series features such as "corresponding position pollutant concentration and change trend" in the aligned sensor readings, and the semantic vector of "spatial feature-time series data" association is generated through geographic space data embedding technology; this association operation can reduce the error interference of single data type (such as satellite image boundary blur and local data deviation of sensor), and finally reduce the error rate of spatial environmental information such as basin pollution range and pollutant concentration in the structured text generated by S100.

[0129] In the embodiment of the present application, in the process of generating the portrait of the agent role, the method further comprises:

[0130] Step 1, pre-injecting role capability evaluation index of Prompt template:

[0131] In the S200 "preset Prompt template parameter injection" link, the role capability evaluation index configuration is added: in the template "ability adaptation" module, three types of indexes such as space governance task proficiency (such as the historical compliance rate of the agent completing the same type of governance task), policy and regulation matching degree (such as the degree of fit between the operation of the agent and the regional governance policy orientation), and cross-regional cooperation response speed (such as the response time of the agent to the collaborative instruction of other regional agents) are supplemented; the index definition and calculation logic (such as task proficiency = historical completion compliance task number / total task number) are embedded in the template parameters, so that when the large language model generates the role image, it can output the ability evaluation results containing the three types of indexes, providing quantitative basis for subsequent image optimization.

[0132] Step 2, role capability evaluation index timing update and image adjustment:

[0133] After S200 "role image verification and associated storage", start the timing update process: set a fixed period (such as every 1 hour), retrieve the regional environmental change information (such as the switching of a certain regional governance task type, the increase of cross-regional cooperation demand) returned by the perception layer in real time, and compare it with the current role capability evaluation index; if the environmental change causes the index deviation (such as the increase of cross-regional cooperation demand makes the "cross-regional cooperation response speed" index not up to standard), dynamically adjust the role image skill weight (such as increasing the weight of "collaborative communication skill") and the space working range boundary (such as expanding the cooperation area coverage); after updating, the new index and the adjusted image are synchronized to the central database to ensure that the role capability is always adapted to the real-time governance scene.

[0134] In the embodiment of the application, in the process of obtaining the corresponding disassembly and distribution result, the method further comprises:

[0135] Step 1, after S300 "Divide-and-Conquer algorithm realizes total task disassembly" and before "subtask and role image semantic mapping", a space-time constraint judgment step is added: based on the geographical coordinate associated region of each subtask (such as the geographical range of a certain street garbage classification supervision subtask and the same street traffic dredging subtask), first compare the geographical boundaries of adjacent subtasks to determine whether there is overlap; then check the execution time window of the overlapping subtasks to confirm whether there is a time conflict. If both "geographical overlap + time conflict" are met, the governance demand urgency level of the subtask (such as the emergency rescue subtask with higher emergency degree than daily patrol) is retrieved, and the subtask with higher emergency degree is preferentially retained for execution according to the original time sequence, and the execution time of the other subtask is delayed to a conflict-free period to avoid interference of the same regional task execution.

[0136] Step 2, in the S300 "semantic mapping of sub-tasks and role profiles and calculation of matching probability" link, a real-time load rate parameter is added: when calculating the matching probability, in addition to the original sub-task query, agent capability, and historical allocation results, the current real-time task load rate of each agent (such as the number of existing unfinished tasks and the proportion of occupied computing resources) is additionally called. The load rate is converted into a weight factor (such as reducing the matching probability weight when the load rate exceeds 80%), which is integrated into the matching probability calculation formula, making the result more consistent with the actual carrying capacity of the agent. For example, if an agent has a high skill matching degree but the real-time load rate is 90%, the final matching probability will be reduced, avoiding task execution delays due to overloading and improving the fit between the matching result and the agent's carrying capacity.

[0137] In the embodiments of the present application, the method further comprises:

[0138] Step 1, the decision layer calls the spatial governance scenario knowledge base:

[0139] Before "integrating agent suggestions" in S400 "decision layer multi-dimensional governance scheme generation and task set determination", the decision layer calls the pre-constructed spatial governance scenario knowledge base. This knowledge base stores historical governance cases (such as past watershed pollution governance and certain urban emergency control cases) and corresponding scheme effect records (such as the execution effectiveness of the scheme in the case and the reason for not meeting expectations). By calling the knowledge base, historical experience is provided for the generation of the current governance scheme, avoiding the design of schemes without basis.

[0140] Step 2, large language model comparison information generates feasibility evaluation report:

[0141] After calling the knowledge base, the decision layer uses a large language model to perform semantic comparison between the real-time spatial information (such as the current regional environmental state and governance task demand) from the perception layer and the historical information of similar scenarios (such as historical cases and scheme effects under similar environmental states) in the knowledge base, analyzes the fit between the current scheme idea and the historical effective scheme, and identifies potential risk points. Based on the comparison results, a scheme feasibility evaluation report is generated, which needs to clearly indicate the scheme pass probability and supporting basis.

[0142] Step 3, scheme pass rate determination and execution layer coordinate verification:

[0143] When the scheme pass rate in the evaluation report is ≥85%, the decision layer generates the final governance scheme based on the agent suggestions; before entering the S400 "execution layer instruction encoding and precise transmission" link, the execution layer needs to perform coordinate verification on the geographic metadata (such as task area latitude and longitude, and key operation point coordinates) in the instructions, compare them with the standard geographic coordinates stored in the central database, correct the coordinate deviation, and ensure that the instructions accurately point to the target governance area, avoiding execution deviation caused by coordinate errors.

[0144] In the embodiments of the present application, in the process of automatically reallocating the roles, the method further comprises:

[0145] Step 1, feedback information classification and scene adaptation weight initial matrix preset:

[0146] After S500 "intelligent agent performs multi-dimensional collection and classification of feedback information", a new classification dimension is added: the feedback information is further subdivided according to the space governance type (ecological monitoring, disaster emergency, urban management and control) and the regional complexity (core urban area, remote mountainous area, etc.), for example, "feedback information of a disaster emergency task in a certain core urban area" and "feedback information of an ecological monitoring task in a certain remote mountainous area"; at the same time, a scene adaptation weight initial matrix is preset, different governance types and regional complexities in the matrix correspond to different weights (for example, the weight of the disaster emergency type is higher than that of the ecological monitoring, and the weight of the core urban area is higher than that of the remote mountainous area), and the matrix replaces the traditional fixed dimension weighting method, so that the feedback analysis is more suitable for the scene characteristics.

[0147] Step 2, feedback information semantic reflection and scenario memory storage and weight correction:

[0148] After S500 "semantic analysis and key feature extraction of feedback information", the analyzed feedback information is subjected to semantic reflection by a large language model, for example, analyzing whether the low role adaptation degree is due to insufficient skills in the disaster emergency scene or poor environmental adaptability in the remote mountainous area, and generating a reflection result; the reflection result is stored as scenario memory and injected into the Prompt of subsequent role generation; at the same time, the weight initial matrix of step 1 is dynamically corrected based on the scenario memory, for example, if the reflection shows that the feedback deviation of the disaster emergency in the remote mountainous area is large for many times, the weight proportion of this scene is increased, and the feedback analysis accuracy is optimized.

[0149] Step 3, role adaptation degree determination and tool evolution process start:

[0150] In the process of S500 "role generation strategy based on feedback feature to update the large language model", an adaptation degree monitoring is added: the role adaptation degree in the feedback information is statistically calculated in real time, and if it is lower than a preset threshold (such as adaptation degree < 60%) for three times in a row, the tool evolution process is automatically started; a tool calling strategy is generated by the large language model combined with the current governance demand (such as "unmanned aerial vehicle aerial photography data rapid analysis" tool is needed for disaster emergency in remote mountainous area), which clearly defines the tool type and the calling trigger condition; at the same time, the skill boundary of the corresponding role image is updated, the operation skill of the new tool is included in the role skill range, and the adaptation ability of the role and the scene is improved.

[0151] In the embodiments of the present application, in addition to the condition trigger and the event trigger, a capability decay trigger is added, and the method further comprises:

[0152] Step 1, capability decay trigger determination and resource threshold dynamic adjustment:

[0153] In S500 "multi-level trigger mechanism deployment and role automatic reassignment execution", a capability attenuation trigger process is added: based on the scenario memory stored in the foregoing, the success rate of the agent is monitored in real time, and if the success rate of the execution of the agent decreases by more than 30% with the task round, it is immediately determined that the capability is attenuated, and the role reassignment is automatically triggered (the updated role generation strategy is called to match the new role); at the same time, the resource threshold trigger adopts a dynamic calculation model, and the threshold is adjusted according to the current governance task urgency (such as relaxing the CPU utilization threshold to 90% for disaster emergency tasks, and keeping 80% for daily tasks), to avoid unnecessary load migration.

[0154] Step 2, rapid initialization of extreme scenario backup agent:

[0155] When the extreme scenario (such as collective capability attenuation of the original role, sudden disaster causing multiple role failures) starts the backup agent, the "historical optimal role-task matching trajectory" (such as the highest adaptation degree of the role skill in the past similar task, the task allocation logic) is extracted from the scenario memory and migrated to the backup agent; then the role parameters (such as adjusting the skill weight, the spatial working range) of the backup agent are fine-tuned in combination with the real-time spatial information (such as the regional environment of the current unfinished task, the subtask type) of the perception layer; at the same time, the pre-stored core role lightweight model snapshot of the edge node is called to directly load and complete the initialization, and the backup agent switching delay is shortened to seconds.

[0156] Step 3, backup agent capability coordination verification:

[0157] After the backup agent starts running, it establishes communication with other node agents through the spatial perception protocol, and carries out capability coordination verification: comparing the skill list (such as environment monitoring, task execution skill) of the backup agent with the demand (such as water quality monitoring, pollution control subtask in a certain area) of the current unfinished subtask, confirming whether the skills completely cover; if there is a skill gap, other node agents are immediately linked to supplement the adaptive capability, to ensure that there is no execution gap in the unfinished subtask.

[0158] The embodiment of the application discloses a multi-agent spatial coordination governance system, which refers to Figure 2 , comprising:

[0159] The multi-source heterogeneous data processing and storage module 001 processes multi-source heterogeneous data composed of satellite images, sensor real-time readings, policy texts and geographic coordinate data, establishes corresponding cross-type semantic association based on geographic spatial data embedding technology, optimizes satellite images through vLLM dynamic memory management, associates regional environment parameters and governance demand of geographic coordinates to generate spatial attribute semantic expression, generates a unified structured text, and transmits the text to the central database through a distributed consistency synchronization mechanism;

[0160] The role image dynamic generation module 002 generates an agent role image including a name, a skill, a knowledge boundary, and a spatial working range based on a preset prompt template based on the spatial environment information in the structured text, the governance target constraint, and the agent capability characteristics when the environment changes or the demand adjusts, and generates the agent role image without relying on fixed rules, wherein the knowledge boundary includes policy and regulation adaptation instructions, and the spatial working range includes associated geographic coordinates.

[0161] The total task decomposition and intelligent distribution module 003 obtains role image and total task information of spatial governance, selects a target computing node based on total task characteristics and role semantic attributes in an edge-center collaborative architecture, splits the total task information into subtasks with dependency relationships by using a Divide-and-Conquer algorithm, establishes semantic mapping between the subtasks and the role image, calculates a matching probability, constructs a reward function according to a scene to optimize a distribution rule, and obtains a corresponding decomposition and distribution result.

[0162] The governance scheme generation and instruction output module 004 generates a corresponding multi-dimensional governance scheme based on the decomposition and distribution result and real-time spatial data in a perception layer-decision layer-execution layer architecture, determines an agent task set, and outputs a governance operation instruction, wherein the perception layer transmits real-time data to the decision layer after the real-time data are processed by an edge node, the decision layer integrates agent suggestions based on a large language model and processes conflicts according to semantic priorities, and the execution layer encodes geographic metadata and instructions based on a spatial perception protocol and transmits the geographic metadata and the instructions to corresponding agents.

[0163] The feedback optimization and multi-level triggering guarantee module 005 obtains execution feedback information corresponding to the agents, updates a role generation strategy and a distribution rule of the large language model based on the execution feedback information, and sets a multi-level triggering mechanism, wherein a sensor data threshold value starting condition trigger and a sudden disaster event trigger are used for automatic role redistribution.

[0164] The embodiment of the application further discloses a multi-agent spatial collaborative governance system, comprising a processor, and a program of the multi-agent spatial collaborative governance method in any one of the above is run in the processor.

[0165] The embodiment of the application further discloses a storage medium, and the program of the multi-agent spatial collaborative governance method in any one of the above is stored in the storage medium.

[0166] Although the embodiments of the application have been shown and described above, it should be understood that the above embodiments are exemplary and cannot be understood as limiting the application, and those skilled in the art can make changes, modifications, replacements, and variations to the above embodiments within the scope of the application.

Claims

1. A multi-agent space collaborative governance method, characterized in that, Comprise: Multi-source heterogeneous data composed of satellite images, sensor real-time readings, policy texts and geographic coordinate data are processed. First, the satellite images are optimized using the vLLM dynamic memory management technology. Memory resources are allocated adaptively according to image resolution, number of bands and regional coverage. Feature extraction speed is improved through pre-compilation and batch processing. Second, based on geospatial data embedding technology, the pixel-level spatial features of satellite images, the location and time sequence value features of sensor real-time readings, the policy provisions and spatial range features of policy texts, and the spatial position features of geographic coordinates are mapped to the same geospatial semantic dimension. Cross-type semantic associations between the four types of data are established through vector similarity calculation. At the same time, taking geographic coordinates as the core link, the corresponding regional environmental parameters and governance needs are associated to generate spatial attribute semantic expressions including coordinate position, environmental state and governance target. Then, the semantic association results and spatial attribute semantic expressions are integrated to generate a unified structured text containing data source type, geographic spatial range, core information content and semantic association identifier. Finally, the unified structured text is transmitted to the central database through the multi-node data verification and incremental synchronization strategy of the edge node and the center node bidirectional checking and the distributed consistency synchronization mechanism. Based on the spatial environmental information, governance target constraints and agent capability characteristics in the structured text, when the environment changes or the demand adjusts, the central server generates an agent role image including name, skill, knowledge boundary and spatial working range based on the pre-set Prompt template using the large language model. The role image does not rely on fixed rules. The knowledge boundary includes policy and regulation adaptation instructions, and the spatial working range includes associated geographic coordinates. The role image and the total task information of spatial governance are obtained. In the edge-center collaborative architecture, the target computing node is selected based on the total task characteristics combined with the semantic attributes of the role. The target node uses the Divide-and-Conquer algorithm to split the total task information into subtasks with dependencies, establishes semantic mapping between the subtasks and the role image, calculates the matching probability, and optimizes the distribution rules based on the scene to obtain the corresponding splitting and distribution results. In the perception layer-decision layer-execution layer architecture, based on the splitting and distribution results combined with real-time spatial data, the corresponding multi-dimensional governance scheme is generated, the agent task set is determined, and the governance operation instruction is output. The perception layer transmits real-time data to the decision layer through the edge node, the decision layer integrates agent suggestions based on the large language model and processes conflicts according to semantic priority, and the execution layer standardizes the geographic metadata and governance operation instruction based on the spatial perception protocol, and then transmits the encoded instruction to the corresponding agent through the exclusive communication link established by the spatial perception protocol. The execution feedback information corresponding to the agent is obtained, and the role generation strategy and distribution rules of the large language model are updated based on the execution feedback information. A multi-level trigger mechanism is established. When the sensor data exceeds the threshold, the conditional trigger is started and the role is automatically redistributed. When a sudden disaster is monitored, the event trigger is started and the role is automatically redistributed.

2. The multi-agent spatial co-governance method of claim 1, wherein, In the process of processing multi-source heterogeneous information, the method further comprises: The real-time sensor readings are first subjected to outlier rejection and time series recording alignment through Kalman filtering algorithm; Before vLLM dynamic memory management optimization, multi-band image fusion technology is used to enhance the spatial environment features of satellite images to improve the accuracy of subsequent spatial environment feature extraction; Based on the enhanced satellite images and aligned sensor readings, semantic association is established to reduce the error rate of spatial environment information in the generated structured text.

3. The multi-agent spatial co-governance method of claim 2, wherein, In the process of generating the agent role image, the method further comprises: The pre-set Prompt template includes role capability evaluation indicators, wherein the role capability evaluation indicators include spatial governance task proficiency, policy and regulation matching degree, and cross-regional collaboration response speed; After the central server large language model generates the role image, it will also combine the real-time regional environment change information transmitted back by the perception layer to update the role capability evaluation indicators once a time, and dynamically adjust the skill weight and spatial working range boundary in the role image.

4. The multi-agent spatial co-governance method of claim 1, wherein, In the process of obtaining the corresponding disassembly and distribution results, the method further comprises: Based on the geographical coordinate associated region of the subtask, a time and space constraint determination step is added; If the geographical ranges of adjacent subtasks overlap and the execution times conflict, the subtask with high governance demand urgency is retained and the execution time sequence of the other subtask is adjusted; When calculating the matching probability, in addition to the comprehensive subtask query, agent capability and historical distribution results, the agent real-time task load rate is also introduced to improve the fitness of the matching probability calculation result to the actual carrying capacity of the agent.

5. The multi-agent spatial co-governance method of claim 1, wherein, The method further comprises: Before integrating the agent suggestions, the decision layer calls the pre-constructed spatial governance scenario knowledge base, which includes historical governance cases and corresponding scheme effect records; By comparing the real-time spatial information with the historical information of similar scenarios in the knowledge base through the large language model, a corresponding scheme feasibility evaluation report is generated; When the pass rate of the scheme in the evaluation report is ≥85%, the corresponding final governance scheme is generated based on the agent suggestions, and before the execution layer outputs the governance operation instructions, the coordinate of the geographical metadata in the instructions is verified.

6. The multi-agent spatial co-governance method of claim 1, wherein, In the process of automatically redistributing the roles, the method further comprises: The feedback information is classified according to spatial governance type and regional complexity, including ecological monitoring, disaster emergency, and urban control, and regional complexity includes core urban area and remote mountainous area; wherein a pre-set scenario adaptation weight initial matrix is used to replace the fixed dimension weighting method; The feedback information is subjected to semantic reflection by the large language model to obtain the corresponding reflection result, which is stored as scenario memory and injected into the subsequent Prompt, and the weight matrix is dynamically corrected based on the scenario memory; if the role adaptation degree of the feedback information is less than the threshold value for three consecutive times, the tool evolution process is automatically started, the large language model generates tool calling strategies based on the governance needs, and the skill boundary of the role image is updated.

7. The multi-agent spatial co-governance method of claim 6, wherein, In addition to conditional triggering and event triggering, a capability decay trigger is added, and the method further comprises: Based on the episodic memory, if the success rate of the agent decreases by more than 30% with the task round, it is determined that the ability is attenuated, and the role reassignment is automatically triggered; and the resource threshold trigger adopts a dynamic calculation model, which adjusts in real time according to the current governance task urgency; When the standby agent is started in an extreme scenario, the historical optimal role-task matching track in the episodic memory is migrated to the standby agent, and the role parameters are fine-tuned in real time combined with the perception layer real-time spatial information; at the same time, the lightweight model snapshot of the core role is pre-stored in the edge node, and the standby agent can directly call the snapshot to complete the initialization when it is started, shortening the switching delay to seconds; After the standby agent runs, it performs ability coordination verification with other node agents through the spatial perception protocol to ensure that the standby agent skill covers the current unfinished subtasks.

8. A multi-agent spatial collaborative governance system, characterized in that, It includes: A multi-source heterogeneous data processing and storage module processes multi-source heterogeneous data composed of satellite images, sensor real-time readings, policy texts, and geographic coordinate data. First, the vLLM dynamic memory management technology is used to optimize satellite images, and memory resources are allocated adaptively according to image resolution, band number, and regional coverage. The feature extraction speed is improved through pre-compilation and batch processing; Based on geographic spatial data embedding technology, pixel-level spatial features of satellite images, location and timing value features of sensor real-time readings, policy provisions and spatial range features of policy texts, and spatial position features of geographic coordinates are mapped to the same geographic spatial semantic dimension, and cross-type semantic associations between the four types of data are established through vector similarity calculation. At the same time, taking geographic coordinates as the core link, the corresponding regional environmental parameters and governance needs are associated to generate spatial attribute semantic expressions including coordinate position, environmental state, and governance target. Then, the semantic association results and spatial attribute semantic expressions are integrated to generate unified structured text containing data source type, geographic spatial range, core information content, and semantic association identifier. Finally, through the multi-node data verification and incremental synchronization strategy of the edge node and the center node, the unified structured text is transmitted to the center database; The role image dynamic generation module generates an agent role image including name, skill, knowledge boundary, and spatial working range based on the spatial environment information, governance target constraints, and agent ability characteristics in the structured text. The image is generated dynamically by the center server large language model based on the preset Prompt template, and does not rely on fixed rules. The knowledge boundary includes policy and regulation adaptation instructions, and the spatial working range includes associated geographic coordinates. The total task decomposition and intelligent distribution module obtains the total task information of the role image and spatial governance, and selects the target calculation node based on the total task characteristics and role semantic attributes in the edge-center collaborative architecture. The target node uses the Divide-and-Conquer algorithm to split the total task information into subtasks with dependency relationships, establishes semantic mapping between the subtasks and the role image, calculates the matching probability, and optimizes the distribution rules based on the scene to obtain the corresponding decomposition and distribution results. The governance scheme generation and instruction output module generates corresponding multi-dimensional governance schemes based on the disassembly and distribution results combined with real-time spatial data in the perception layer-decision layer-execution layer architecture, determines the agent task set and outputs the governance operation instructions, wherein the perception layer is transmitted to the decision layer after the real-time data is processed by the edge node, the decision layer integrates the agent suggestions based on the large language model and processes the conflicts according to the semantic priority, the execution layer standardizes the coding of the geographic metadata and the governance operation instructions based on the spatial perception protocol, and then the coded instructions are accurately transmitted to the corresponding agent through the exclusive communication link established by the spatial perception protocol; The feedback optimization and multi-level trigger guarantee module obtains the execution feedback information corresponding to the agent, updates the role generation strategy and distribution rule of the large language model based on the execution feedback information, and sets up a multi-level trigger mechanism, wherein when the sensor data is over the threshold value, the conditional trigger is started and the roles are automatically redistributed; when a sudden disaster is monitored, the event trigger is started and the roles are automatically redistributed.

9. A multi-agent spatial collaborative governance system, characterized in that, A processor, wherein the processor runs a program of the multi-agent spatial collaborative governance method according to any one of claims 1-7.

10. A storage medium, characterized by A storage device, wherein the storage device stores a program of the multi-agent spatial collaborative governance method according to any one of claims 1-7.

Images